Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

MiniMax Code (mcode)

MiniMax's open-source terminal agent, built on Pi's libraries, with a cloud classifier deciding tool approvals by default.

Reviewed by ReviewRamble editors · Updated

View on GitHub Open source

Share this page:XLinkedInFacebookRedditWhatsAppEmail

Our rating

3.5/5

MiniMax Code is a careful, well-documented terminal agent with sensible telemetry defaults, but it is young, closed to outside contributions and leans on MiniMax's cloud even to decide what it may run.

What is MiniMax Code (mcode)?

MiniMax Code is MiniMax's coding agent. The terminal part, `mcode`, is open source under MIT in MiniMax-AI/minimax-code and offers an interactive TUI, a headless `mcode exec` mode for scripts and CI, an ACP mode for editors such as Zed, and a TypeScript SDK. The companion desktop app for macOS and Windows shares the runtime but its source is not published.

Under the hood it uses Earendil's Pi packages (pi-agent-core, pi-ai and pi-coding-agent) for the agent loop, model protocols and terminal, with MiniMax's own permissions, sandbox, plugins, subagents and account services on top. It signs in to a MiniMax account to use the M Plan subscription, or runs without a login on any OpenAI- or Anthropic-compatible endpoint.

Is MiniMax Code (mcode) worth it?

MiniMax Code is MiniMax's answer to Claude Code, and the open-source part is the terminal agent rather than the desktop app. It covers the expected ground: reading and editing files, running commands and tests, plan mode, subagents, plugins and skills, MCP, and sessions you can resume, fork, rewind and export. A lightweight mode sends a short conversational prompt with no tools, so simple questions cost fewer tokens.

It is built on Pi's libraries, the same foundation as omp, which makes it a sibling of Pi more than a new design. The repository was created in June 2026, has 2,270 stars and released v0.7.2 on 31 October 2026, with commits as recent as 17 November. The README describes the repository as a source preview and says pull requests are accepted only from collaborators.

Telemetry is better handled than in most rivals. Usage events, runtime metrics and error diagnostics are each off by default with separate opt-ins, and MCODE_DISABLE_TELEMETRY or DO_NOT_TRACK switches every channel off. Login, model requests and update checks are separate from those switches.

The catch: the default permission mode is Auto, in which a MiniMax cloud service classifies each tool call using conversation context; if the classifier is unavailable, mcode falls back to asking you. Ask and Full access are the other modes. The sandbox exists, with read-only, workspace-write and delete-guard levels, but ships switched off with full filesystem and network access. MiniMax does not publish M Plan prices on the pages we could read; tiers are sold as Go, Explore and Build at 1, 3 and 7.5 times Go's quota.

Our take: with your own key and Ask mode, mcode is a tidy Pi-based agent. On MiniMax's account, more of your session touches MiniMax's servers than the open source alone suggests.

Where it earns its place

  • A developer on a MiniMax M Plan

    Use the plan's quota for coding without managing an API key

    Run `mcode login --region global`, then `mcode` in the project; /status shows the account. Image and video models count only against the weekly limit.

  • A QA engineer at a mid-sized software firm

    Fix failing tests overnight in CI

    Call `mcode exec` with a prompt such as finding and fixing a failing test, inside a container, with the sandbox set to workspace_write and network denied in config.yaml.

  • A privacy-conscious contractor working on client code

    Keep prompts and approval decisions away from the vendor

    Add a self-hosted endpoint with `mcode provider add`, skip the MiniMax login and switch the permission mode to Ask with Alt+M or /permission.

Before you commit

Setup
`curl -fsSL https://filecdn.minimax.chat/public/install.sh | bash` (installs into ~/.minimax-code, no sudo) or npm with Node.js 22.19+; then `mcode login` or `mcode provider add` for your own key.
Integrations
MiniMax models and M Plan; any openai-completions, openai-responses or anthropic-messages endpoint; MCP and managed connectors; plugins and skills; ACP for Zed; TypeScript SDK `@minimax-ai/code-sdk`.
Data and privacy
Prompts and code go to the chosen provider. In Auto mode tool calls and conversation context go to MiniMax's permission classifier. User data lives in ~/.minimax. Usage, metrics and diagnostics telemetry are each off by default.
Cost at scale
No paid edition of the agent. M Plan quotas scale 1, 3 and 7.5 times from Go to Build; prices were not shown on the official pages we read. On the API, M3.1 Flash costs $0.15 input, $0.60 output and $0.015 cached per 1M tokens.
Limits
Sandbox off by default; Alpine and musl not supported by the one-command installer; outside pull requests are not accepted.

What’s good

  • Every telemetry channel off by default
  • TUI, headless, ACP and SDK from one MIT codebase
  • Works without a MiniMax login on your own endpoint
  • Optional sandbox with read-only and workspace-write levels

What’s not

  • Default Auto mode sends tool calls to a cloud classifier
  • Sandbox off by default
  • Pull requests limited to collaborators; source preview
  • M Plan prices not shown on the official pages we read

Best for

Developers already on MiniMax models who want the vendor's own terminal agent, or Pi users who want MiniMax's plugins and subagents ready-made.

MiniMax Code (mcode) alternatives

MiniMax Code (mcode): common questions

Is MiniMax Code (mcode) worth it?
MiniMax Code is a careful, well-documented terminal agent with sensible telemetry defaults, but it is young, closed to outside contributions and leans on MiniMax's cloud even to decide what it may run. We rate it 3.5 out of 5.
How much does MiniMax Code (mcode) cost?
Free and open source (MIT). MiniMax models through an M Plan subscription (Go, Explore, Build) or the API, for example M3.1 Flash at $0.15 input and $0.60 output per 1M tokens; or your own key.
What are the downsides of MiniMax Code (mcode)?
Default Auto mode sends tool calls to a cloud classifier; sandbox off by default; pull requests limited to collaborators; source preview; m Plan prices not shown on the official pages we read.
Who is MiniMax Code (mcode) best for?
Developers already on MiniMax models who want the vendor's own terminal agent, or Pi users who want MiniMax's plugins and subagents ready-made.
How much does MiniMax Code (mcode) cost for a team or in production?
No paid edition of the agent. M Plan quotas scale 1, 3 and 7.5 times from Go to Build; prices were not shown on the official pages we read. On the API, M3.1 Flash costs $0.15 input, $0.60 output and $0.015 cached per 1M tokens.
Is MiniMax Code (mcode) safe for company data?
Prompts and code go to the chosen provider. In Auto mode tool calls and conversation context go to MiniMax's permission classifier. User data lives in ~/.minimax. Usage, metrics and diagnostics telemetry are each off by default.
What are the best MiniMax Code (mcode) alternatives?
The closest alternatives we have reviewed are Pi, Qwen Code and OpenCode.

Answers checked Nov 20, 2026.